Prediction of extinction ratio's temperature compensation table using neural network

Abstract

The conventional way of modeling extinction ratio’s (ER) temperature compensation table of a transceiver module results to high manufacturing testing time, thus gives an issue in manufacturing line which is low UPH (Units per hour). The conventional way is through manual temperature cycling and through an algorithm which is step search. This uses an expensive time with low UPH. In an ER temperature compensation table, several TOSA and module parametric values affect it. Each of these parameters were studied and used to feed the network. This work aims on determining the best parameters that will produce AC bias based on relevant AC properties, developing an MLP ANN model that utilizes and identifies parameters in order to predict AC bias value which will be used in generating ER temperature compensation table, and lastly, modeling an artificial neural network that predicts ER temperature compensation table to boost up UPH. Several experiments were performed to select the best parameters to produce AC bias based on relevant AC properties and these are all TOSA data and Module_ER_at_80, which includes a total of 26 parameters. In addition to this, optimum UPH is obtained using these parameters at 2.44. And the MLP ANN model with 23 number of neurons in hidden layer was developed to obtain the highest possible neural network performance which is having 2.44UPH, +/-3 DAC counts distribution, ῀8.7 MSE, and ῀0.93 r-square.